CLARA: A Context-Aware Drug Recommendation Framework Using Multi-Level Integration of Large Language Models and Graph Neural Networks

Personalized drug recommendation remains a formidable challenge in clinical decision support, dealing with intricate patient data, sparse labeled outcomes, and limited interpretability of existing models. In this context, we introduce CLARA, a unified framework for drug recommendation that integrates semantic patient modeling, graph-based learning, and natural language explanations powered by Large language models (LLMs). Using structured and unstructured electronic health records from the MIMIC-III dataset, we construct a heterogeneous graph that links patients, diagnoses, procedures, medications, and clinical notes. Contextual embeddings, derived via a SentenceTransformer model, capture the depth of patient profiles, while a Graph Attention Network (GAT) predicts top-k medication classes through relational reasoning. To improve interpretability, an LLM–driven module delivers clear, human-readable justifications for each recommendation. Evaluations reveal that CLARA exceeds established baselines in both predictive accuracy and clinical transparency, as validated on a MIMIC-III subset. By integrating LLMs for context extraction and explanation, our approach advances safe, personalized, and explainable AI-driven medical recommendations.